The top things worth knowing about in AI today.
Anthropic's most capable models, Claude Fable 5 and Mythos 5, remain unavailable ten days after a US Commerce Department export-control order pulled them on 12 June over security concerns. The free trial that had included Fable 5 on paid plans ends today with the models still down, even as Anthropic says it expects access to return within days. It is the first time leading US models have been restricted like export-controlled technology, leaving businesses that began building on them in limbo.
Read moreSalesforce agreed on 15 June to acquire Fin, the customer service agent formerly known as Intercom, for about $3.6 billion, its largest purchase of an agentic customer-service company. Fin's AI agent resolves queries end to end across chat, email, WhatsApp, SMS and phone, and will be folded into Salesforce's Agentforce platform. The deal is another sign that automated front-line support is shifting from experiment to default.
Read moreNoam Shazeer, who co-wrote the 2017 paper that underpins modern AI and most recently co-led Google's Gemini, said on 17 June he is joining OpenAI to lead architecture research. Google paid roughly $2.7 billion in 2024 to bring him back from Character.AI; he stayed less than two years. The move shows how hard the top labs are competing for the small group of people who shape how frontier models are built.
Read moreChinese lab Zhipu (Z.ai) released GLM-5.2, an open-weight model with a 1-million-token context window under an MIT licence that allows free commercial use and self-hosting. Independent tests rank it the strongest open-weight coding model, close behind Claude Opus 4.8 and matching or beating GPT-5.5 on long coding tasks at roughly a sixth of the cost. For teams watching their AI bills, capable open models are steadily narrowing the gap with the priciest closed options.
Read moreSecurity researchers disclosed "agentjacking," which plants hidden instructions in fake bug reports so AI coding tools such as Claude Code and Cursor run an attacker's commands while investigating them. Tests against more than 100 organisations succeeded 85% of the time, and at least 2,388 companies were found exposed through public Sentry error-tracking credentials. Any team that lets an AI agent read outside data and run commands should check what those agents are permitted to execute.
Read moreFrance's Mistral is in talks to raise about 3 billion euros (roughly $3.5 billion) at a valuation near 20 billion euros, which would nearly double its worth from a round less than a year ago. That would make the three-year-old company one of Europe's most valuable private tech firms and the clearest counterweight to the US labs. Terms are early and could still change.
Read moreFrom 15 June, every non-corporate Commonwealth entity must keep an internal register of its AI use cases, each with a named accountable owner, under the Digital Transformation Agency's responsible-AI policy. Mandatory AI training for Australian Public Service staff also became a requirement, and all 94 in-scope agencies have published public AI transparency statements. Further obligations, including impact assessments before deployment and incident reporting, begin in December.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.